A two-stage deep learning approach for accurate day-ahead electricity price forecasting

Participants in the energy market are at greater risk of making decisions due to the nonlinear and volatile characteristics of electricity prices. Accurate short-term electricity price forecasting (EPF) is essential to ensure improved resource allocation, grid stability and enable market participants to manage their decisions efficiently. This study proposes a novel two-stage forecasting framework for day-ahead EPF using time series decomposition methods and hybrid deep learning algorithms. In the first stage, features related to EPF at the next time step are predicted. In this stage, the highest-frequency component extracted via Empirical Mode Decomposition (EMD) is further decomposed using Variational Mode Decomposition (VMD) so as to better capture rapid fluctuations and improve the overall prediction accuracy. Moreover, a decentralized deep learning architecture is designed in which Gated Recurrent Unit (GRU) networks are employed for high-frequency components, while Long Short-term Memory (LSTM) networks are used for the remaining components. In the second stage, EPF is generated using a hybrid LSTM and GRU structure, which incorporates both features estimated in the first stage and historical electricity price data. Finally, hyperparameters of the deep learning models are optimized using Bayesian Optimization to enhance performance. To validate the proposed framework, real market data from the DK1 region of Denmark is used. The proposed hybrid prediction framework is evaluated against both machine learning methods and deep learning-based architectures. Experimental results demonstrate that the proposed method achieves approximately 27.15 % lower RMSE compared to traditional machine learning models, and around 28.24 % lower RMSE compared to LSTM-based models.